Stubble Burning and Punjab
Read more to find out why and how we built Parali Alert, what it does and why it matters.
Every year, as the paddy harvest season ends in Punjab, a familiar problem returns to the fields. Farmers face a narrow window to clear their fields before the next crop cycle, and for many, burning the leftover straw remains one of the fastest and cheapest options. What happens in these fields, however, does not stay in the fields. Smoke travels across districts and states, contributing to deteriorating air quality and exposing millions of people to harmful pollutants. As a punjabi boy who lived in Delhi for most of his life, I can go on and on about the effect of stubble burning not just in Punjab but in other states and how the air that we breathe becomes poisonous.
The Problem:
Paddy cultivation leaves behind substantial quantities of crop residue. After harvesting, farmers must prepare their fields for the next sowing season, often under considerable time and financial pressure. In the absence of accessible machinery, affordable alternatives, or sufficient time, burning the residue can appear to be the most practical solution.
The resulting fires release particulate matter, carbon monoxide, and other pollutants into the atmosphere. Under unfavourable meteorological conditions, these emissions can contribute to severe regional air pollution.
Traditional satellite-based fire monitoring systems play an important role in identifying active fires. However, detecting a fire after ignition is fundamentally different from preventing one. By the time a thermal sensor identifies a burning field, the opportunity for pre-emptive intervention may already have passed.
Agricultural authorities need more than a map of where fires are burning. They need actionable intelligence about where fires are likely to occur next.
Where Parali Alert Fits In:
Parali Alert approaches stubble burning as a predictive geospatial problem.
Instead of relying exclusively on active-fire detections, the system analyses satellite observations, weather conditions, and ground-level operational requirements to identify administrative units where crop residue burning may be imminent.
The critical period is the short interval between harvesting and residue burning. During this window, authorities can potentially deploy residue-management machinery, coordinate field inspections, and contact farmers before ignition occurs.
The objective is to transform satellite data into operational decisions: which areas need attention, when intervention is most useful, and where available resources should be sent first.
Parali Alert and Land Imagery
Parali Alert combines multiple Earth-observation data sources, each providing a different view of agricultural activity.
1. Sentinel-2: Identifying Harvest Events
Optical satellite imagery provides information about vegetation and surface characteristics. Parali Alert uses the Normalized Difference Vegetation Index (NDVI) to monitor changes in vegetation across agricultural land.
A significant decline in NDVI can indicate that a crop has been harvested. By comparing observations across time, the system can identify potential harvest events and estimate how long a field has been in the post-harvest stage.
This temporal information matters because the likelihood of residue burning is not constant throughout the harvesting cycle.
2. Estimating Crop Residue with Spectral Indices
Harvesting does not eliminate the material left on the soil surface. The amount and condition of this residue influence subsequent management decisions.
Parali Alert uses the Normalized Difference Tillage Index (NDTI), calculated from Sentinel-2 short-wave infrared bands, as an additional spectral indicator of surface conditions.
Combined with harvest timing and other observations, this information helps characterize potential residue-management requirements.
Importantly, spectral indices are indirect measurements rather than definitive measurements of straw mass. Their interpretation requires appropriate calibration and validation against field observations.
3. Sentinel-1: Observing Through Cloud Cover
Optical imagery has an important limitation: clouds can obstruct observations during critical periods.
Synthetic Aperture Radar (SAR) provides a complementary source of information. Sentinel-1 transmits radar signals and measures the energy scattered back from the Earth's surface, allowing observations to be collected independently of sunlight and through cloud cover.
Parali Alert incorporates radar backscatter observations to support monitoring when optical data are unavailable or unreliable.
This combination of optical and radar data creates a more robust framework for observing changing agricultural landscapes.
Satellite data alone cannot explain whether a field will burn. Weather, residue conditions, harvesting schedules, and the availability of alternative residue-management equipment all influence the decision.
Parali Alert therefore combines several factors in a deterministic risk engine:
- Harvest likelihood: Signals associated with recent harvesting.
- Residue load: Spectral indicators of crop residue on the ground.
- Dryness factor: Soil moisture and temperature conditions.
- Weather favourability: Meteorological conditions relevant to burning and dispersion.
- Ground demand: Reported requirements for residue-management equipment.
These weights form a transparent scoring framework. Rather than producing an unexplained prediction, the system can expose the factors contributing to an area's calculated priority.
The initial score is then adjusted according to the intervention window and potential downwind impact. Areas approaching the critical post-harvest period can receive greater priority, while wind direction and speed help identify locations whose emissions could affect population centres.
Wind as a Signal
A fire's consequences depend not only on where it starts but also on where its emissions travel.
Parali Alert incorporates meteorological forecasts from Open-Meteo to calculate wind vectors and assess potential downwind transport. This allows the system to distinguish between agricultural units with similar burning risk but different potential consequences for nearby population centres.
The result is a broader approach to prioritization. Authorities can consider both the probability of a burning event and the potential regional impact when deciding where to direct limited personnel and machinery.
Wind-based transport estimates remain approximations: actual smoke dispersion also depends on atmospheric stability, boundary-layer dynamics, plume height, and other meteorological factors. Nevertheless, even a simplified transport model can provide useful additional context for operational planning when appropriately validated.
Where AWS comes into play:
Parali Alert was built from the ground up to demonstrate how modern, serverless cloud engineering can solve one of South Asia’s most catastrophic environmental and public health crises. Rather than maintaining expensive, idle server clusters, our entire end-to-end telemetry, geospatial analysis, and field-dispatch pipeline runs natively on Amazon Web Services (AWS).
Memory Saved with AWS Open Data & S3 HTTP Range Requests:
Traditional remote sensing workflows download entire Sentinel-2 .SAFE satellite archives (~500 MB per tile). By leveraging the Sentinel-2 Cloud-Optimized GeoTIFF (COG) dataset hosted on the Registry of Open Data on AWS (via Element84 on S3), our custom geospatial ingestion engine queries only target administrative unit bounding boxes using HTTP Range requests (Range: bytes=X-Y).
We pull only the exact 100x100 pixel window covering the target village or agricultural block.
This slashes network transfer from 500 MB to ~65 KB per band, a 99.98% reduction in data bandwidth and memory footprint powered directly by AWS S3 infrastructure.
Serverless Analytical Engine with AWS Lambda:
Processing multi-spectral imagery across entire agricultural belts usually demands dedicated cluster hardware. We replaced persistent servers with two focused AWS Lambda functions running on Python 3.11 with custom x86_64 layers containing pre-compiled NumPy and Shapely binaries.
The scheduled assessment function ingests fresh optical data, radar backscatter, and meteorological wind models. It computes Delta NDVI vegetative drops, residue indices, and desiccation curves across all 27 pilot units in Punjab in just 4.6 seconds.
The companion API function wraps our FastAPI backend using the Mangum ASGI adapter. Paired with Amazon API Gateway HTTP API v2, it serves live risk queues, unit GeoJSON files, and tractor allocations with sub-150ms cold starts and native CORS support.
Pre-Dawn Automation with Amazon EventBridge:
Decisions to burn crop residue happen during early morning field inspections. To put actionable data in front of agricultural officers before their workday starts, Amazon EventBridge executes our assessment function automatically at 05:30 AM IST (00:00 UTC) every day.
The rule triggers without manual intervention, guarantees timely execution across all districts, and shuts down compute instances as soon as the calculations finish.
Immutable Regional Archive in Amazon S3:
Every daily risk assessment creates a permanent, versioned record stored in our Amazon S3 Data Lake bucket.
These structured JSON snapshots give administrative supervisors a verifiable audit trail of stubble accumulation, farmer requests, and estimated emission reductions. Field dashboards can also instantly pull past snapshots for comparative time-series analysis without triggering repetitive calculations.
Immediate Field Mobilization with Amazon SNS:
Finding at-risk fields is only useful if local officers can intervene before fires start. When an administrative unit hits critical status, our Lambda engine publishes an alert directly to an Amazon Simple Notification Service (SNS) topic.
Amazon SNS broadcasts high-priority SMS alerts containing unit names, risk percentages, and machine recommendations to field officers and Custom Hiring Centers. This communication loop alerts response teams 36 to 48 hours before ignition takes place, giving operators enough time to dispatch balers and seeders.
Edge Acceleration with Amazon CloudFront and S3 Hosting:
Our mission control interface combines MapLibre GL vector boundaries with Deck.gl 3D wind particle layers. The compiled single-page application is hosted as a static website on Amazon S3 and delivered globally through Amazon CloudFront.
CloudFront provides SSL encryption, edge caching, and low latency for regional control rooms, ensuring that interactive maps and dispatch tables load smoothly on mobile tablets and desktop monitors alike.
Custom Hiring Centers and Machine Allocation
Predicting a fire is only half the battle. If a farmer does not have access to alternatives, telling them not to burn will not solve anything. The bottleneck in Punjab is rarely farmer willingness; it is machine logistics during a frantic two-week window.
Parali Alert integrates directly with village-level Custom Hiring Centers (CHCs). These centers house specialized residue-management equipment:
- Happy Seeders and Super Seeders: Machines that sow wheat directly into standing paddy stubble without needing to clear or burn it.
- Straw Balers and Rakes: Equipment that compacts straw into dense bales for biomass power plants and industrial boilers.
- Mulchers and Shredders: Implements that chop straw into organic mulch to enrich soil nutrients.
When Parali Alert flags a high-priority block, the system does not just send a warning. It checks local machinery inventories and calculates utilization rates across nearby CHCs. Field officers receive a clear action plan detailing which machines are available and where to route them before farmers run out of time.
Backtesting
In geospatial and disaster management, predictive models often look impressive until they are tested against real-world timelines. To ensure our system was operationally sound, we ran a zero-leakage temporal backtest validated against multi-year historical harvest data across Punjab.
We separated our evaluation strictly by time: model parameters were calibrated on October data and tested on the peak burning period between late October and mid-November. No future fire locations were leaked into past predictions.
The results highlight a fundamental shift in operational capability:
- Operational Lead Time: Traditional thermal sensors detect fires roughly 4 hours after ignition has already started. Parali Alert identifies critical intervention windows an average of 37.2 hours before ignition, giving administration a 41-hour operational advantage.
- Precision at Top Ten Percent: Our multi-spectral engine achieved 74.1% precision in identifying top-priority intervention zones, representing a 2.3x improvement over standard reactive heuristics.
- Recall: The system captured 82.3% of eventual fire events within a 48-hour pre-harvest window.
The Impact:
Stubble burning is not just a localized agricultural issue; it is a massive driver of regional air quality degradation and climate change. To quantify the real-world value of proactive intervention, Parali Alert incorporates emission formulas from the Indian Council of Agricultural Research (ICAR Tier-2 guidelines).
For every ton of paddy straw diverted from open burning into in-situ seeding or ex-situ baling:
- Avoided Carbon Dioxide: 1.46 tons of CO2
- Avoided Particulate Matter: 3.0 kilograms of PM2.5
- Avoided Black Carbon: 0.5 kilograms of toxic soot
By modeling tractor allocations across our 27 pilot units, intervening across just 14,000 hectares of cropland prevents over 120,000 tons of greenhouse gases and particulate matter from entering the Indo-Gangetic airshed.
What Lies Ahead:
While our pilot focuses on critical districts in Punjab including Sangrur, Ludhiana, Bathinda, and Tarn Taran, the architecture was engineered to scale. Because every component runs serverlessly on AWS, extending the platform across Haryana, Western Uttar Pradesh, and the wider Indo-Gangetic Plains requires zero changes to server infrastructure.
Our next milestone is integrating conversational voice bots in Punjabi and Hindi through Bhashini. This will allow village sarpanches and individual farmers to request equipment directly via voice notes on WhatsApp, completing the chain from orbital satellites to the farmer in the field.
Try It Live:
Parali Alert is fully deployed and accessible today. You can explore the live 3D mission control console, inspect real-time risk scores across Punjab, and examine wind dispersion vectors directly:
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article